Safe arrays and pointers for C

Transcription

1 Safe arrays and pointers for C John Nagle Discussion draft July, 2012 Introduction Buffer overflows continue to plague C and C++ programs. This is a draft proposal for dealing with that problem. The basis of this proposal is to define means for expressing the size of arrays in C. C already has fixed-size arrays with useful semantics. In this proposal, the existing syntax for fixed-size arrays is generalized to allow known-size arrays, where the size of the array is known at variable initialization time. With relatively minor and compatible changes to C, the most troublesome causes of program crashes and security vulnerabilities can be dealt with. In any useful program, each array has a size known to the programmer. In C, there is currently no way to consistently express that size in the language. This proposal adds that capability. Scope An ISO document on common programming language vulnerabilities, Information Technology Programming Languages Guidance to Avoiding Vulnerabilities in Programming Languages through Language Selection and Use, ISO/IEC TR 24772, provides a standard taxonomy of flaws in programming language designs. The specific flaws in C addressed by this proposal are: 6.8 String Termination 6.9 Buffer Boundary Violation (Buffer Overflow) 6.10 Unchecked Array Indexing 6.11 Unchecked Array Copying 6.12 Pointer Casting and Pointer Type Changes 6.13 Pointer Arithmetic 6.14 Null Pointer Dereference 6.40 Type-breaking Reinterpretation of Data These are the errors most closely associated with program crashes and low-level security vulnerabilities. Not addressed are: 6.15 Dangling Reference to Heap 6.41 Memory Leak 07/20/12 Safe arrays and pointers for C - DRAFT 1

2 Preventing errors of those types requires automatic memory management or extensive static analysis. Those flaws are difficult to fix without unacceptably large changes to the C language. Informally, the problem with adding garbage collection to C is that C is the language in which one writes the garbage collector. Summary of language changes The language changes required are minor. Relaxation of existing restrictions Expressions allowed in most array size declarations. (This is the key enhancement.) Features imported from C++11 auto References New reserved words force_cast New predefined functions lengthof(array) array_slice(array, start, end) void array_init(variable, value) void array_zero(variable) New pragma #pragma strict New restrictions (non-strict mode) None on existing code other than avoidance of the reserved word force_cast. New restrictions (strict mode) Conversion from a pointer to an array to a reference to an array of known length requires an explicit force_cast, to indicate that the programmer has commanded an unsafe operation. Conversion from anything which can be NULL to a reference implies a run-time assert check. Unions may not contain pointer or reference types, and all parts of a union must have the same length. 07/20/12 Safe arrays and pointers for C - DRAFT 2

3 To be used for pointer arithmetic, a pointer variable must be initialized at declaration to an element of an array. The only allowed values for the pointer variable used in pointer arithmetic are ones which point to an element of the array to which it was originally assigned. (Pointing to an element one past the end is allowed, for backwards compatibility. Rationale Goals No change to run-time representation of data. (no descriptors or fat pointers ) Minimal modifications to the C language. Define strict and non-strict forms of C modules; existing C99 code can be compiled as non-strict modules and strict modules can interoperate with non-strict modules. If all modules are of strict form, and no unsafe overrides are used, programs should be protected against the flaws listed above. Hidden memory management mechanisms (garbage collection, reference counting) should not be required. Previous work Microsoft source code annotation language (SAL) Cyclone (AT&T research language) SCC, the Safe C compiler. Ccured MemSafe Incompatible dialects and related languages Fixed-size arrays in C and C++ The basis of this proposal is to define means for expressing the size of arrays in C. Most arrays in C and C++ are passed around as pointers, with no size information. There is, however, an exception fixed sized arrays. C and C++ support fixed-sized arrays. Fixed-size arrays are typed objects of known length. They can be defined as types, and, in C++, references to fixed size arrays are fully supported. int arr100[100]; arr100& arr100r; // an array of 100 ints // ref to array of 100 ints This is valid C++ today. C++ already supports the concept of a reference to an array. (Arrays of references are prohibited by the C++ standard, but references to arrays are not.) 07/20/12 Safe arrays and pointers for C - DRAFT 3

4 Where the compiler knows the size of an array, we can use that information. It can be accessed via sizeof. This is convenient, but only useful for arrays of fixed size. C99 dynamically sized arrays C99 added support for dynamically sized temporary arrays. This feature is often used in numerical code, where it is common to call a function which needs a temporary array for the duration of the function call. float mathfn(const float in[ ], size_t n) { float workarr[n]; // work array of length n } return(result); This was the first place in C where an array could be resized at compile time with the compiler aware of the size. Known-size arrays. The existing syntax and semantics of fixed-size arrays, C99 dynamically sized arrays, and C++ references can, with slight extensions, allow the support of known-sized arrays. The only syntactical change required is to allow expressions in a few places where, at present, only constants are allowed. It is proposed to allow expressions in array sizes in type declarations in the following places: local variable and typedef declarations function parameter declarations casts within initialization expressions structure definitions (under certain restrictions) When a known-sized array is defined or passed as a parameter, its size is the value of the dimension expression evaluated at initialization time. This allows a function such as the following: void copybyref(int (&a)[n], const int (&b)[n], size_t n) { for (int i = 0; i < lengthof(a); i++) { a[i] = b[i]; } } This is a size-safe copy, intended to be used as follows: int a0[100], a1[100]; copybyref(a1, a0); // defined with a type declaration // copy array of known size 07/20/12 Safe arrays and pointers for C - DRAFT 4

5 int b0[50], b1[50]; copybyref(b0, b1, 50); copybyref(b0, a0, 100); // smaller arrays // copy array of a different size // size mismatch compile time error The last copybyref call is an error, because the array sizes do not match. For a constant n this can be caught at compile time. For a variable n, it has to be a run time check. The compiler must check, at a call to copybyref : assert(lengthof(a) == n); assert(lengthof(b) == n); If the programmer calls copybyref as follows: copybyref(b0, a0, lengthof(b0)); The checks become assert(lengthof(b0) == lengthof(b0)); assert(lengthof(a0) == lengthof(b0)); The first check is optimized out as an identity. The second check still requires a run-time check. But only one check per function call is required. This is far cheaper than general subscript checking. Here, a run-time check is required only because two arrays must match. In many common cases, especially some of the classic causes of buffer overflows, there is no additional overhead. For example, the classic UNIX read call would now be expressed as: int read(int fd, char(&buf)[n], size_t n); The required check is If called with assert(lengthof(buf) == n); int buf[512];... int stat = read(fd, buf, lengthof(buf)); the check required is assert(lengthof(buf) == lengthof(buf)); which is optimized out as an identity. Worth noting is that the function declaration of read above is, in non-strict mode, compatible with existing code which passes buf as a pointer to a char. Why references? A key part of this proposal is the use of C++ style references in C. As shown above, references to 07/20/12 Safe arrays and pointers for C - DRAFT 5

6 arrays can already contain size information about the array. Pointers to arrays in C refer to the type of the first element, with no array length information. Changing the semantics of pointers in C would break existing code. Adding references to the C language will not break anything. C++ already has both references and pointers. The relationship between the two is well understood and not a cause of trouble, so this can be done with confidence. Strict mode The enhancements to C syntax and semantics described here are backwards compatible with existing C code. Beyond this is strict mode which disallows some unsafe constructs and operations. Strict mode requires known size arrays in situations where overflow is possible. Strict mode is defined on a persource-file basis. Strict and non-strict code can interoperate. New rules enforced in strict mode: Conversion from a pointer to an array to a reference to an array of known length requires an explicit force_cast, to indicate that the programmer has commanded an unsafe operation. Conversion from anything which can be NULL to a reference implies a run-time assert check. Unions cannot contain pointers or references, and all alternatives of a union must have the same size. Programs written in strict mode should be protected against the list of memory safety errors covered in the scope section above. This should eliminate most buffer overflow problems. Strict mode for a compilation unit is indicated by or _Pragma("strict") #pragma strict Implications of known-size arrays With known-sized arrays and strict mode introduced, their use and consequences can be examined. How array size information is stored There are no array descriptors involved. The programmer tells the compiler the size of the array as an expression, and that expression is evaluated at the point in the program where the relevant declaration appears. int asize = 100; // size wanted int atab[asize]; // declare known-size array asize = 200; // this does not resize the array printf( Length of atab is %d\n, lengthof(atab)); // prints 100 In the above worst case example, the compiler must generate a temporary is needed to store the size of the array, because the inputs to the expression defining its size change. This is usually unnecessary. 07/20/12 Safe arrays and pointers for C - DRAFT 6

7 Reference declarations C99 already allows variable declarations to be interspersed with executable code. References must be initialized in C++, and this restriction would be maintained with references in C. So references will generally be initialized near the point of first use. We are now allowing expressions in array sizes in declarations, and that expression must be evaluated where the declaration appears. int b[100];... int (&rb)[lengthof(b)] = b; // array of 100 ints // ref to array b Assignments to references are subject to type checking. Where the reference on the left hand side has a known length determined at run time, this may imply a run-time assert check on array size. This is considered a subscript check, as defined below. Casts The syntax of casts follows that for declarations. As with reference declarations, the size must be evaluated at the point of the cast. Casts with expressions in array sizes are allowed in initialization expressions. Ordinary casts must be memory-safe in strict mode. Unsafe casts must be expressed using some specific syntax, as with reinterpret_cast in C++. Unfortunately, the syntax of reinterpret_cast uses the corner-bracket template notation of C++, and is inappropriate for C. We need some syntax to indicate that a cast is forced in violation of normal type rules. This, like most syntax issues is a politically touchy problem. A solution should be Consistent with the expectation of C programmers. Backwards compatible with existing code Findable by simple searches to enable easy code auditing. One option is a special storage class, used only in casts. A phrase such as forced_cast might be appropriate: int val = 1234; char* ptr = (forced_cast char*)val; // UNSAFE Even in strict mode, this would be required only when forcing a conversion which could impact memory safety. So such casts would be quite rare. A few casts can generate run-time checks. Conversion of a pointer to a reference must generate an assert check. Such checks are enabled when subscript checking is enabled. Dynamic allocation ( malloc and its friends) This is one of the more difficult problems to handle safely. Yet, as it is frequently a source of program bugs of the buffer overflow variety, it it necessary to do so. The problems which must be addressed include: The classical return type of malloc, void *, cannot express size information. 07/20/12 Safe arrays and pointers for C - DRAFT 7

8 Because malloc can return NULL, it is not permissible to directly generate a reference from memory allocation. Nor can allocation failure be handled via an exception. Pointers to references are not allowed. Structures which contain references must be initialized. The syntax for allocating memory must be concise. Backwards compatibility with existing code must be maintained in non-strict mode. All those constraints, taken together, create a tough problem. A few options: Allocation functions which raise a signal on failure. This impacts program structure, but provides the cleanest syntax. char (&p)[n] = MALLOCARRAY(char, n); with a suitable macro, is a clean, concise syntax, but there's no way to handle the error condition in in-line code. It's worth providing this for routine applications which do not need sophisticated out-of-memory handling. Return a pointer to a reference to the object, or null. With some encapsulation in a macro, we can insure that the necessary casting is type-safe. #define MALLOC(T) (T*)malloc(sizeof(T)) // safe allocator If we use this to allocate an array of known size, the result is a pointer to the array as an object, or NULL. size_t n = 100; // desired array size... typedef int tab[n]; tab* p = MALLOC(tab); if (p == NULL) return(false); tab& tab1 = *p; // get ref from ref to pointer... tab1[10] = 1; // array of known size n // alloc a "tab" // handle allocation failure // it's a valid array This is type-safe and size-safe. Note that there are no casts other than inside the define. Initialization of dynamically allocated space In C, malloc normally provides memory space containing old semi-random data. For most data types, this is not a safety problem. However, for pointer safety, we must insure that allocated items with pointers or references are initialized. Zeroing new memory space is not an option for references, and because we're encouraging the use of references in place of pointers, this has to be done right. So, when a new struct or array or structs is allocated, and the struct type contains a pointer or reference, it must be initialized from an initial value of that type. 07/20/12 Safe arrays and pointers for C - DRAFT 8

9 Syntax for this has yet to be defined. Something like struct item { size_t len; char (&) msg[len]; }; const char[] msg1 = Hello, world ; const itemproto = { sizeof(msg1), msg1); item*p = MALLOC(item, itemproto); suggests itself. Underneath that macro, a built-in generic function is required for the initialization operation. This is a safe replacement for the classic bcopy. void array_init(variable, value) This simply copies value into variable, with type checking. The variable parameter can be an array of known or fixed size. This is needed because array assignment isn't allowed, and assignment to an array reference assigns the reference, not the content. In C++, we could write this as a template function, but in C, any generic function must be a built-in. As a convenience, a safe replacement for the common bzero is provided: void array_zero(variable) This simply sets the array to binary zero. It is a compile-time error to use this on an array of structures containing references. Associating length with an array pointer in a structure We might now want a structure which carries both the array and its size. The pointer-based form is typedef str { size_t len, char* data}; New reference form: typedef str { size_t len, char (&data)[len]); str s1 = {n, MALLOC(char, n) }; // structure initialization The key point here is that char (&data)[len] is a reference to a string of known length. The length, len, is an element in the same structure. This ties the length of the array to a value which can be found from the structure. This tie is a property of the type declaration. The general rule for such declarations is that the dimension expression is evaluated in the context of the struct. Fields of the struct may appear in the expression. As a convenience, lengthof(s1.data) is then well defined. 07/20/12 Safe arrays and pointers for C - DRAFT 9

10 This idiom is only valid for structures initialized all at once through structure initialization. A struct with an array as the last component A special case is typedef struct item { size_t itemlength; char itemvalue[itemlength]; }; This is permitted. The size of the object is calculated based on the value of itemlength. Such structs should be initialized as a unit, with an initialization expression. In strict mode, assignment to the length is not permitted after initialization. Impact on other language features The basic proposal has now been described. The impact of the proposal on various features of the C language is next. Unions In strict mode, elements of unions are restricted to non-pointer and non-reference types, and all alternatives must be the same size. In most existing code, unions are used primarily for data being passed through networks, files, and message passing systems. Such data, since it comes from external sources, seldom contains pointers. Adding language support for safe discriminated variant records, as was done in Cyclone, seems excessive for the infrequency with which it would be used. In strict mode, the only way to override pointer type safety is through the explicit use of force_cast. This make it easy to search and audit programs for pointer type problems. Multidimensional arrays Multidimensional arrays of known size are supported. This simply generalizes the existing rules for multidimensional fixed-size arrays. size_t order = 3; float tab[order][order]; // square array of order N. As with arrays of fixed size, these are not arrays of pointers, but dense arrays of elements. This will be a great convenience in numerical work. There is also a performance gain. On modern CPUs, multiplying by row size to compute a row offset is faster than accessing a table of pointers. Pointer arithmetic Pointer arithmetic is a standard idiom of C programming. Although it would be safer to eliminate it from the language, that isn't feasible for an update to C. 07/20/12 Safe arrays and pointers for C - DRAFT 10

11 The most common uses of pointer arithmetic use local variables which are bound to a single array throughout their life. Such pointers behave much like C++ iterators. So we can allow pointer arithmetic under much the same restrictions applied to iterators. These are: The pointer variable must be initialized at its declaration to an element of an array. The only allowed values for the pointer variable are ones which point to an element of the array to which it was originally assigned. (Pointing to an element one past the end is allowed, for backwards compatibility. Pointers with such properties are safe for pointer arithmetic purposes, provided there is range checking. Range checking for such pointers is possible without fat pointers, because the array to which they are bound is known at compile time. Unsafe pointer arithmetic is allowed only through force_cast. String constants String constants have type char (&)[length] or wchar_t(&)[length]. These types will convert to char* or wchar_t* for compatibility with non-strict code. The sizeof and lengthof functions can be applied and count the usual trailing null. Null-terminated strings are a problem. A workable solution is to allow read access to const arrays without subscript checking. This allows reading junk, but not overwriting other variables. Standard library safety Library function can be categorized as follows: Can be made fully compatible and safe with arrays of known size read, write, fread, fwrite, etc. Unsafe and already deprecated gets, sprintf, etc. These would be unavailable in strict mode. Safe for write, but not for read printf, snprintf, etc. Permitted in strict mode on the grounds that the impact of disallowing them is too high. Unsafe sscanf, which stores into memory based on format strings. Unavailable in strict mode. Library issues are beyond the scope of this document, but appear to be manageable. Subscript checking With enough information available to perform subscript checking, implementations may offer it as an option. The question is what to do when a run time error is detected. The available options are to take the action defined for an assert failure, or to raise a signal. That decision can be left to the implementation. It is strongly recommended that implementations which do subscript checking optimize it. As a minimum, the following is suggested: common subexpression optimization should be applied to subscript checks within loops. Within for loops, checks should be hoisted to the beginning of the loop and done once per 07/20/12 Safe arrays and pointers for C - DRAFT 11

12 loop if at all possible. Hoisted checks which resolve to constant expressions or identities should be evaluated at compile time, and reported as errors (if failing) or optimize out (if succeeding). The array must not have a scope narrower than that of the pointer variable, so that the array cannot go out of scope while the pointer variable is still in scope. The goal of the above is to optimize out most, if not all, subscript checks in inner loops of common matrix calculations. It is explicitly permitted to detect and act on a subscript error as soon as it becomes inevitable. In other words, if an implementation can detect that a subscript error will occur on some iteration of a loop, and there is no way to exit the loop before reaching that iteration, the subscript error can be reported at entry to the loop. A potential exit from the loop via break or return must inhibit this optimization. A potential exit via longjmp or exit, which might be hidden in a called function, would not be considered to inhibit this optimization. Convenience features With the language aware of array length, some convenience features can be added to make programming easier. These are all optional from the programmer's perspective, and need not be retrofitted to existing code. auto storage type as in C++ 20xx. auto, from C++11, is added to C. The motivation for this is the long declarations required for references to complex arrays. lengthof C has a built-in sizeof function. It's common to define lengthof, returning the number of elements of an array, as a macro. It is proposed to make lengthof a standard. Both sizeof and lengthof are defined for arrays of known size. This makes the size of an array immediately and reliably available to code. Iteration over an array Almost all languages defined in the last 20 years (with the notable exception of JavaScript) have syntax for iterating over all the elements of a collection. For arrays of known size, the compiler now has enough information to provide such a construct. The syntax of this may be controversial, and needs discussion. Some options include for (auto x in arr) { } for (auto x forall arr) { } Making in a reserved word impacts existing code, but yields the nicest syntax. Using this feature with auto leads to concise loop statements which are not error prone. Suggestions in this area are welcome. 07/20/12 Safe arrays and pointers for C - DRAFT 12

13 Subarray syntax At times, it is desirable to talk about a subarray of an array. The existing C syntax for this is char arr[100] char p[] = &arr[20]; This tends to be error-prone, as the size of the subarray is rather vague. So it is proposed to borrow the slice operation from Javascript: char arr[100] char (&p)[80] = array_slice(arr, 20, 100); Following the Javascript and Python convention, the limits of slice are the start index and the end index + 1. Negative, overlapping, and out of bounds indices are errors and should be detected if subscript checking is enabled. Another option is to borrow Python's subarray syntax: Char arr[100]; char (&p)[80] = arr[20:100]; // reference to a subarray of arr This is concise, but more controversial, as it involves new syntax. It has the advantage that subarray references could be allowed on the left side of an assignment. Conclusion Alternatives Making C a memory safe language without unduly changing the language is difficult. This proposal is intended as a direction for standard C, and as such, must be quite conservative in the changes proposed. Annotation systems, such as Microsoft's SAL, can be effective if used. Because they're not integrated into the language, they tend to make programs more verbose. Thus, they're rarely used without strong managerial pressure. Cyclone has a reasonable set of extensions, but strays far enough from C that it's a different language. Cyclone offers three types of pointers - normal, never-null, and fat. New symbols and? ) are used to designate the different types of pointers. Fat pointers imply hidden run-time machinery. This is a valid approach, but it is not C. SCC, the Safe C compiler, is claimed to be able to detect all pointer and access errors at run time. However, it required using an elaborate data structure for every pointer and substantial run-time overhead. C++ addresses safety by trying to hide the unsafe constructs of C under a layer of standard templates. In practice, the templated constructs tend to leak raw C pointers, and buffer overflows in C++ applications are still common. Other variants such as Objective-C and Java are useful languages, but they are not C. 07/20/12 Safe arrays and pointers for C - DRAFT 13

14 Limitations This approach still doesn't solve the dangling-pointer problem. There's no good way to do that without more memory management machinery behind the scenes. The goal here was to avoid adding hidden machinery. 07/20/12 Safe arrays and pointers for C - DRAFT 14

15 Appendix 1 Examples From the UNIX/POSIX API: Original function declarations: New form: int read(int fd, char* buf, size_t n); int read(int fd, char (&)buf[n], size_t n); This is size-safe when called from new code, and compatible with old code that passes a char* as a parameter. From Numerical Recipes in C, Gauss-Jordan elimination: Original function declaration: void gaussj(float **a, int n, float **b, int m); New form: void gaussj(float (a&)[n][n], int n, float (&b)[n][m], int m); In the original, the array bounds have to be explained in comments. (More to follow) 07/20/12 Safe arrays and pointers for C - DRAFT 15